Rotationally Invariant Descriptors using Intensity Order ... - IEEE Xplore
Rotationally Invariant Descriptors using Intensity Order ... - IEEE Xplore Rotationally Invariant Descriptors using Intensity Order ... - IEEE Xplore
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IEEE TRANSACTION ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 18 order based partitions can encode ordinal information into descriptor. Therefore, the proposed descriptors could have higher discriminative ability. (3) Since intensity orders are invariant to monotonic intensity changes, our proposed descrip- tors provide a higher degree of illumination invariance, not merely to the linear illumination change. Thus they can deal with large illumination changes, especially for MRRID, it has much better results than MROGH and other evaluated descriptors when matching images exhibit large illumination changes (see Section V-C.2). This is because for MRRID, not only its feature pooling scheme is based on intensity orders, its local feature is also based on the relative intensity relationship of sample points. (4) The proposed descriptors are constructed on the basis of multiple support regions, further enhancing their discriminative ability. By utilizing multiple support regions, it also avoids the problem of selecting an optimal region size to construct descriptor for a detected interest region to some extent. A. Parameters Evaluation V. EXPERIMENTS There are several parameters in the proposed descriptors: the number of spatial partitions k, the number of support regions N, the number of orientation bins d, and the number of binary codes m. As listed in Table I, MROGH and MRRID share two parameters: the number of spatial partitions and the number of support regions, while the number of orientation bins is needed in MROGH and the number of binary codes is needed in MRRID. In order to evaluate their influences on the performance of the proposed descriptors, we conducted image matching experiments on 142 pairs of images 2 with different parameter settings as listed in Tab. I. These 142 image pairs are mainly selected from the dataset of zoom and rotation transformations. Note that they do not contain image pairs in the standard Oxford dataset [44] because those image pairs are used for the descriptors evaluation in the later stage. Fig. 9 shows the average recall vs. average 1-precision curves of MROGH and MRRID with different parameter settings. The definition of a correct match and a correspondence is the same as [22] which is determined with overlap error [9]. The matching strategy used here is the nearest 2 They are real images downloaded from [43]. The ground truth homography is supplied along with the image pair. November 26, 2011 DRAFT
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IEEE TRANSACTION ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 19 MROGH MRRID TABLE I PARAMETERS OF OUR PROPOSED DESCRIPTORS denotation parameter settings description k 4,6,8 number of spatial partitions d 4,8 number of orientation bins N 1,2,3,4 number of support regions k 4,6,8 number of spatial partitions m 3,4 number of binary codes N 1,2,3,4 number of support regions neighbor distance ratio [22]. In the remaining experiments, we used the same definitions of a match, a correct match and a correspondence. It can be seen from Fig. 9 that the performances of MROGH and MRRID are improved when the number of support regions is increased. When a single support region is used, MROGH performs the best with the parameter setting of ’d=8,k=8’, closely followed by the setting of ’d=8,k=6’. While MRRID performs the best with the parameter setting of ’m=4,k=8’, followed by the setting of ’m=4,k=4’. Taking into consideration of performance and complexity (dimension), we use the parameter setting of ’d=8,k=6,N=4’ for MROGH and ’m=4,k=4,N=4’ for MRRID in the remaining experiments. Thus the dimensionality is 192 for MROGH and 256 for MRRID. recall 1 0.8 0.6 d=4,k=4,N=1(16) 0.4 d=4,k=6,N=1(24) d=4,k=8,N=1(32) 0.2 d=8,k=4,N=1(32) d=8,k=6,N=1(48) d=8,k=8,N=1(64) 0 0 0.1 0.2 0.3 0.4 1−precision 0.5 0.6 recall 1 0.8 0.6 0.4 0.2 d=8,k=6,N=1(48) d=8,k=6,N=2(96) d=8,k=6,N=3(144) d=8,k=6,N=4(192) 0 0 0.1 0.2 0.3 1−precision 0.4 0.5 (a) Gradient-Based Descriptor (MROGH) recall 1 0.8 0.6 0.4 m=3,k=4,N=1(32) m=3,k=6,N=1(48) m=3,k=8,N=1(64) 0.2 m=4,k=4,N=1(64) m=4,k=6,N=1(96) m=4,k=8,N=1(128) 0 0 0.1 0.2 0.3 0.4 1−precision 0.5 0.6 recall 1 0.8 0.6 0.4 0.2 m=4,k=4,N=1(64) m=4,k=4,N=2(128) m=4,k=4,N=3(192) m=4,k=4,N=4(256) 0 0 0.1 0.2 0.3 1−precision 0.4 0.5 (b) Intensity-Based Descriptor (MRRID) Fig. 9. The average performance of the proposed descriptors with different parameter settings. B. Multi-Support Regions vs. Single Support Region This experiment aims to show the superiority of using multi-support regions over a single support region. We used the same dataset as in the experiment of parameters evaluation (Sec- tion V-A). Since four support regions are used in our method, for each of the proposed descrip- November 26, 2011 DRAFT
- Page 1 and 2: This article has been accepted for
- Page 3 and 4: This article has been accepted for
- Page 5 and 6: This article has been accepted for
- Page 7 and 8: This article has been accepted for
- Page 9 and 10: This article has been accepted for
- Page 11 and 12: This article has been accepted for
- Page 13 and 14: This article has been accepted for
- Page 15 and 16: This article has been accepted for
- Page 17: This article has been accepted for
- Page 21 and 22: This article has been accepted for
- Page 23 and 24: This article has been accepted for
- Page 25 and 26: This article has been accepted for
- Page 27 and 28: This article has been accepted for
- Page 29 and 30: This article has been accepted for
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.<br />
<strong>IEEE</strong> TRANSACTION ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 18<br />
order based partitions can encode ordinal information into descriptor. Therefore, the proposed<br />
descriptors could have higher discriminative ability.<br />
(3) Since intensity orders are invariant to monotonic intensity changes, our proposed descrip-<br />
tors provide a higher degree of illumination invariance, not merely to the linear illumination<br />
change. Thus they can deal with large illumination changes, especially for MRRID, it has much<br />
better results than MROGH and other evaluated descriptors when matching images exhibit large<br />
illumination changes (see Section V-C.2). This is because for MRRID, not only its feature<br />
pooling scheme is based on intensity orders, its local feature is also based on the relative intensity<br />
relationship of sample points.<br />
(4) The proposed descriptors are constructed on the basis of multiple support regions, further<br />
enhancing their discriminative ability. By utilizing multiple support regions, it also avoids the<br />
problem of selecting an optimal region size to construct descriptor for a detected interest region<br />
to some extent.<br />
A. Parameters Evaluation<br />
V. EXPERIMENTS<br />
There are several parameters in the proposed descriptors: the number of spatial partitions k,<br />
the number of support regions N, the number of orientation bins d, and the number of binary<br />
codes m. As listed in Table I, MROGH and MRRID share two parameters: the number of<br />
spatial partitions and the number of support regions, while the number of orientation bins is<br />
needed in MROGH and the number of binary codes is needed in MRRID. In order to evaluate<br />
their influences on the performance of the proposed descriptors, we conducted image matching<br />
experiments on 142 pairs of images 2 with different parameter settings as listed in Tab. I. These<br />
142 image pairs are mainly selected from the dataset of zoom and rotation transformations. Note<br />
that they do not contain image pairs in the standard Oxford dataset [44] because those image<br />
pairs are used for the descriptors evaluation in the later stage.<br />
Fig. 9 shows the average recall vs. average 1-precision curves of MROGH and MRRID with<br />
different parameter settings. The definition of a correct match and a correspondence is the same<br />
as [22] which is determined with overlap error [9]. The matching strategy used here is the nearest<br />
2 They are real images downloaded from [43]. The ground truth homography is supplied along with the image pair.<br />
November 26, 2011 DRAFT
Change language